Qwen runtime completion control

All 32 requests completed with 32 output tokens each in three vLLM 0.29.0 runs and two vLLM 0.30.0 runs on one H100 NVL.

Exact test

• Model: Qwen/Qwen3-0.6B, revision c1899de289a04d12100db370d81485cdf75e47ca, bfloat16, no quantization.
• Upstream random offline throughput benchmark: 128 input tokens, 32 output tokens, 32 measured requests, 8 warmups.
• One physical NVIDIA H100 NVL, driver 580.126.09; tensor parallelism 1, maximum model length 1024, memory utilization 0.5, eager execution, prefix caching disabled.
• The same input-token hash, warmup and measured fixture hashes, model files, GPU UUID, driver and common flags were verified across the five arms.
• Official release images changed dependent library versions as well as vLLM. This is a comparison of these recorded release environments; no effect is attributed exclusively to vLLM source.

What supports the length claim

The executed observer rejects unfinished requests, any count other than 32, multiple outputs, or an output length other than 32 tokens. The retained runtime bundle receipt binds its hash, and all five observations record PASS. The output token sequences were not exported, so their contents cannot be audited retrospectively. Matching output length does not demonstrate correct answers.

A/A and scope

The three baseline runs passed the same categorical completion and length checks. They do not establish output-content identity. The timing observations do not establish a calibrated speed comparison.

Downloadable record

Versioned JSON report - exact images, parameters, times, model-file hashes, package differences and evidence-file hashes. This historical record applies only to the named input and release pair.

• Baseline package inventory
• Candidate package inventory
• Pinned model receipt
• Actual common flags
• Fresh primary-source receipts

Primary sources

• Baseline upstream benchmark
• Candidate upstream benchmark
• Pinned model revision